episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
because there is only one past but there are
https://karpathy.ai/lexicap/0011-large.html#00:51:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
many possible futures and so a reinforcement
https://karpathy.ai/lexicap/0011-large.html#00:51:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
learning system which is trying to maximize its
https://karpathy.ai/lexicap/0011-large.html#00:51:19.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
future expected reward and doesn't know yet which
https://karpathy.ai/lexicap/0011-large.html#00:51:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of these many possible futures should I select
https://karpathy.ai/lexicap/0011-large.html#00:51:26.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
given this one single past is facing problems
https://karpathy.ai/lexicap/0011-large.html#00:51:29.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that the LSTM by itself cannot solve.
https://karpathy.ai/lexicap/0011-large.html#00:51:33.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So the LSTM is good for coming up with a compact
https://karpathy.ai/lexicap/0011-large.html#00:51:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
representation of the history and observations
https://karpathy.ai/lexicap/0011-large.html#00:51:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and actions so far but now how do you plan in an
https://karpathy.ai/lexicap/0011-large.html#00:51:44.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
efficient and good way among all these, how do
https://karpathy.ai/lexicap/0011-large.html#00:51:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you select one of these many possible action
https://karpathy.ai/lexicap/0011-large.html#00:51:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
sequences that a reinforcement learning system
https://karpathy.ai/lexicap/0011-large.html#00:51:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
has to consider to maximize reward in this
https://karpathy.ai/lexicap/0011-large.html#00:52:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
unknown future?
https://karpathy.ai/lexicap/0011-large.html#00:52:04.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
We have this basic setup where you have one
https://karpathy.ai/lexicap/0011-large.html#00:52:06.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
recurrent network which gets in the video and
https://karpathy.ai/lexicap/0011-large.html#00:52:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the speech and whatever and it's executing
https://karpathy.ai/lexicap/0011-large.html#00:52:14.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
actions and it's trying to maximize reward so
https://karpathy.ai/lexicap/0011-large.html#00:52:17.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
there is no teacher who tells it what to do at
https://karpathy.ai/lexicap/0011-large.html#00:52:20.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which point in time.
https://karpathy.ai/lexicap/0011-large.html#00:52:23.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And then there's the other network which is
https://karpathy.ai/lexicap/0011-large.html#00:52:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
just predicting what's going to happen if I do
https://karpathy.ai/lexicap/0011-large.html#00:52:29.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that and that and that could be an LSTM network
https://karpathy.ai/lexicap/0011-large.html#00:52:32.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and it learns to look back all the way to make
https://karpathy.ai/lexicap/0011-large.html#00:52:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
better predictions of the next time step.
https://karpathy.ai/lexicap/0011-large.html#00:52:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So essentially although it's predicting only the
https://karpathy.ai/lexicap/0011-large.html#00:52:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
next time step it is motivated to learn to put
https://karpathy.ai/lexicap/0011-large.html#00:52:44.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
into memory something that happened maybe a
https://karpathy.ai/lexicap/0011-large.html#00:52:48.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
million steps ago because it's important to
https://karpathy.ai/lexicap/0011-large.html#00:52:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
memorize that if you want to predict that at the
https://karpathy.ai/lexicap/0011-large.html#00:52:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
next time step, the next event.
https://karpathy.ai/lexicap/0011-large.html#00:52:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Now how can a model of the world like that, a
https://karpathy.ai/lexicap/0011-large.html#00:52:59.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
predictive model of the world be used by the
https://karpathy.ai/lexicap/0011-large.html#00:53:03.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
first guy?
https://karpathy.ai/lexicap/0011-large.html#00:53:06.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Let's call it the controller and the model, the
https://karpathy.ai/lexicap/0011-large.html#00:53:07.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
controller and the model.
https://karpathy.ai/lexicap/0011-large.html#00:53:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
How can the model be used by the controller to
https://karpathy.ai/lexicap/0011-large.html#00:53:12.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
efficiently select among these many possible
https://karpathy.ai/lexicap/0011-large.html#00:53:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
futures?
https://karpathy.ai/lexicap/0011-large.html#00:53:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
The naive way we had about 30 years ago was
https://karpathy.ai/lexicap/0011-large.html#00:53:19.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
let's just use the model of the world as a stand
https://karpathy.ai/lexicap/0011-large.html#00:53:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in, as a simulation of the world and millisecond
https://karpathy.ai/lexicap/0011-large.html#00:53:26.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
by millisecond we plan the future and that means
https://karpathy.ai/lexicap/0011-large.html#00:53:30.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
we have to roll it out really in detail and it
https://karpathy.ai/lexicap/0011-large.html#00:53:33.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
will work only if the model is really good and
https://karpathy.ai/lexicap/0011-large.html#00:53:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
it will still be inefficient because we have to
https://karpathy.ai/lexicap/0011-large.html#00:53:39.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
look at all these possible futures and there are
https://karpathy.ai/lexicap/0011-large.html#00:53:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
so many of them.
https://karpathy.ai/lexicap/0011-large.html#00:53:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So instead what we do now since 2015 in our CM
https://karpathy.ai/lexicap/0011-large.html#00:53:46.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
systems, controller model systems, we give the
https://karpathy.ai/lexicap/0011-large.html#00:53:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
controller the opportunity to learn by itself how
https://karpathy.ai/lexicap/0011-large.html#00:53:52.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to use the potentially relevant parts of the M,
https://karpathy.ai/lexicap/0011-large.html#00:53:56.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of the model network to solve new problems more
https://karpathy.ai/lexicap/0011-large.html#00:54:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
quickly.
https://karpathy.ai/lexicap/0011-large.html#00:54:04.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And if it wants to, it can learn to ignore the M
https://karpathy.ai/lexicap/0011-large.html#00:54:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and sometimes it's a good idea to ignore the M
https://karpathy.ai/lexicap/0011-large.html#00:54:09.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
because it's really bad, it's a bad predictor in
https://karpathy.ai/lexicap/0011-large.html#00:54:12.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
this particular situation of life where the
https://karpathy.ai/lexicap/0011-large.html#00:54:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
controller is currently trying to maximize reward.
https://karpathy.ai/lexicap/0011-large.html#00:54:19.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
However, it can also learn to address and exploit
https://karpathy.ai/lexicap/0011-large.html#00:54:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
some of the subprograms that came about in the
https://karpathy.ai/lexicap/0011-large.html#00:54:26.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
model network through compressing the data by
https://karpathy.ai/lexicap/0011-large.html#00:54:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
predicting it.
https://karpathy.ai/lexicap/0011-large.html#00:54:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it now has an opportunity to reuse that code,
https://karpathy.ai/lexicap/0011-large.html#00:54:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the algorithmic information in the model network
https://karpathy.ai/lexicap/0011-large.html#00:54:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to reduce its own search space such that it can
https://karpathy.ai/lexicap/0011-large.html#00:54:44.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
solve a new problem more quickly than without the
https://karpathy.ai/lexicap/0011-large.html#00:54:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
model.
https://karpathy.ai/lexicap/0011-large.html#00:54:52.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Compression.
https://karpathy.ai/lexicap/0011-large.html#00:54:53.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So you're ultimately optimistic and excited about
https://karpathy.ai/lexicap/0011-large.html#00:54:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the power of RL, of reinforcement learning in the
https://karpathy.ai/lexicap/0011-large.html#00:54:59.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
context of real systems.
https://karpathy.ai/lexicap/0011-large.html#00:55:03.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Absolutely, yeah.
https://karpathy.ai/lexicap/0011-large.html#00:55:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So you see RL as a potential having a huge impact
https://karpathy.ai/lexicap/0011-large.html#00:55:06.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
beyond just sort of the M part is often developed on
https://karpathy.ai/lexicap/0011-large.html#00:55:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
supervised learning methods.
https://karpathy.ai/lexicap/0011-large.html#00:55:16.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
You see RL as a for problems of self driving cars
https://karpathy.ai/lexicap/0011-large.html#00:55:19.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
or any kind of applied cyber robotics.
https://karpathy.ai/lexicap/0011-large.html#00:55:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
That's the correct interesting direction for
https://karpathy.ai/lexicap/0011-large.html#00:55:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
research in your view?
https://karpathy.ai/lexicap/0011-large.html#00:55:32.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
I do think so.
https://karpathy.ai/lexicap/0011-large.html#00:55:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
We have a company called Nasence which has applied
https://karpathy.ai/lexicap/0011-large.html#00:55:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
reinforcement learning to little Audis which learn
https://karpathy.ai/lexicap/0011-large.html#00:55:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to park without a teacher.
https://karpathy.ai/lexicap/0011-large.html#00:55:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
The same principles were used of course.
https://karpathy.ai/lexicap/0011-large.html#00:55:47.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So these little Audis, they are small, maybe like
https://karpathy.ai/lexicap/0011-large.html#00:55:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that, so much smaller than the real Audis.
https://karpathy.ai/lexicap/0011-large.html#00:55:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
But they have all the sensors that you find in the
https://karpathy.ai/lexicap/0011-large.html#00:55:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
real Audis.
https://karpathy.ai/lexicap/0011-large.html#00:56:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
You find the cameras, the LIDAR sensors.
https://karpathy.ai/lexicap/0011-large.html#00:56:01.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
They go up to 120 kilometers an hour if they want
https://karpathy.ai/lexicap/0011-large.html#00:56:03.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to.
https://karpathy.ai/lexicap/0011-large.html#00:56:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And they have pain sensors basically and they don't
https://karpathy.ai/lexicap/0011-large.html#00:56:09.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
want to bump against obstacles and other Audis and
https://karpathy.ai/lexicap/0011-large.html#00:56:13.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
so they must learn like little babies to park.
https://karpathy.ai/lexicap/0011-large.html#00:56:17.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Take the raw vision input and translate that into
https://karpathy.ai/lexicap/0011-large.html#00:56:21.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
actions that lead to successful parking behavior
https://karpathy.ai/lexicap/0011-large.html#00:56:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which is a rewarding thing.
https://karpathy.ai/lexicap/0011-large.html#00:56:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And yes, they learn that.
https://karpathy.ai/lexicap/0011-large.html#00:56:30.040